Agent observability and trace graph

Agent Observability helps developers answer three production questions: what the agent attempted, what KeyRunner allowed, and what the downstream API returned.

Find a trace

Open Observe in AI Workspace. Search by trace ID, agent, tool, policy, or error, then filter by agent and outcome. Summary metrics show success rate, average tool latency, model tokens, and estimated cost when the runtime reports those fields.

Token or cost values may be zero or unavailable when the upstream model integration does not report usage.

Read trace details

Selecting a trace opens its execution details. Depending on the recorded activity, the timeline can include:

• Agent invocation
• One or more model calls
• Policy evaluation
• Governed tool calls
• Credential context without exposing the credential value
• API responses and status
• Duration, token usage, cost, and errors

Open the trace graph

Select Open trace graph for a visual execution map. The graph lays out the execution horizontally and groups repeated model, tool, and response stages for readability. You can pan, zoom, fit the graph to the viewport, and drag nodes while investigating a large trace.

Use the graph to identify slow tools, repeated calls, policy boundaries, downstream failures, and the point at which an agent execution diverged from the expected path.

Data availability

Observability reflects telemetry received by KeyRunner. Missing tokens, model names, cost, or response details normally mean the client or runtime did not report that field; it does not necessarily indicate a failed execution.